Message ID | alpine.LSU.2.20.1705311603540.20726@zhemvz.fhfr.qr |
---|---|
State | New |
Headers | show |
Richard Biener <rguenther@suse.de> writes: > So I've come back to PR66313 and found a solution to the tailrecursion > missed optimization when fixing the factoring folding to use an unsigned > type when we're not sure of overflow. > > The folding part is identical to my last try from 2015, the tailrecursion > part makes us handle intermittent stmts that were introduced by foldings > that "clobber" our quest walking the single-use chain of stmts between > the call and the return (and failing at all stmts that are not part > of said chain). A simple solution is to move the stmts that are not > part of the chain and that we can move before the call. That handles > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > > Richard. > > 2017-05-31 Richard Biener <rguenther@suse.de> > > PR middle-end/66313 > * fold-const.c (fold_plusminus_mult_expr): If the factored > factor may be zero use a wrapping type for the inner operation. > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap > and handle moved defs. > (process_assignment): Properly guard the unary op case. Return a > tri-state indicating that moving the stmt before the call may allow > to continue. Pass through to_move. > (find_tail_calls): Handle moving unrelated defs before > the call. > > * c-c++-common/ubsan/pr66313.c: New testcase. > * gcc.dg/tree-ssa/loop-15.c: Adjust. > > Index: gcc/fold-const.c > =================================================================== > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 > *************** fold_plusminus_mult_expr (location_t loc > *** 6916,6925 **** > } > same = NULL_TREE; > > ! if (operand_equal_p (arg01, arg11, 0)) > ! same = arg01, alt0 = arg00, alt1 = arg10; > ! else if (operand_equal_p (arg00, arg10, 0)) > same = arg00, alt0 = arg01, alt1 = arg11; > else if (operand_equal_p (arg00, arg11, 0)) > same = arg00, alt0 = arg01, alt1 = arg10; > else if (operand_equal_p (arg01, arg10, 0)) > --- 6916,6926 ---- > } > same = NULL_TREE; > > ! /* Prefer factoring a common non-constant. */ > ! if (operand_equal_p (arg00, arg10, 0)) > same = arg00, alt0 = arg01, alt1 = arg11; > + else if (operand_equal_p (arg01, arg11, 0)) > + same = arg01, alt0 = arg00, alt1 = arg10; > else if (operand_equal_p (arg00, arg11, 0)) > same = arg00, alt0 = arg01, alt1 = arg10; > else if (operand_equal_p (arg01, arg10, 0)) > *************** fold_plusminus_mult_expr (location_t loc > *** 6974,6987 **** > } > } > > ! if (same) > return fold_build2_loc (loc, MULT_EXPR, type, > fold_build2_loc (loc, code, type, > fold_convert_loc (loc, type, alt0), > fold_convert_loc (loc, type, alt1)), > fold_convert_loc (loc, type, same)); > > ! return NULL_TREE; > } > > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > --- 6975,7010 ---- > } > } > > ! if (!same) > ! return NULL_TREE; > ! > ! if (! INTEGRAL_TYPE_P (type) > ! || TYPE_OVERFLOW_WRAPS (type) > ! /* We are neither factoring zero nor minus one. */ > ! || TREE_CODE (same) == INTEGER_CST) > return fold_build2_loc (loc, MULT_EXPR, type, > fold_build2_loc (loc, code, type, > fold_convert_loc (loc, type, alt0), > fold_convert_loc (loc, type, alt1)), > fold_convert_loc (loc, type, same)); > > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise > ! same may be minus one and thus the multiplication may overflow. Perform > ! the operations in an unsigned type. */ > ! tree utype = unsigned_type_for (type); > ! tree tem = fold_build2_loc (loc, code, utype, > ! fold_convert_loc (loc, utype, alt0), > ! fold_convert_loc (loc, utype, alt1)); > ! /* If the sum evaluated to a constant that is not -INF the multiplication > ! cannot overflow. */ > ! if (TREE_CODE (tem) == INTEGER_CST > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) > ! return fold_build2_loc (loc, MULT_EXPR, type, > ! fold_convert (type, tem), same); > ! > ! return fold_convert_loc (loc, type, > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, > ! fold_convert_loc (loc, utype, same))); > } > > /* Subroutine of native_encode_expr. Encode the INTEGER_CST Sorry if you already know, but this part means that we can no longer vectorise: int f (int *x, int b1, int b2, int b3) { int foo = 0; for (int i1 = 0; i1 < b1; ++i1) for (int i2 = 0; i2 < b2; ++i2) for (int i3 = 0; i3 < b3; ++i3) foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; return foo; } We now convert all the arithmetic in the [...] to unsigned int and reassociate it so that the "- 1" is applied last. We then assume that the overflow in this subtraction is well-defined and that the &x[...] might not be linear for the inner loop. Thanks, Richard
On Tue, 13 Jun 2017, Richard Sandiford wrote: > Richard Biener <rguenther@suse.de> writes: > > So I've come back to PR66313 and found a solution to the tailrecursion > > missed optimization when fixing the factoring folding to use an unsigned > > type when we're not sure of overflow. > > > > The folding part is identical to my last try from 2015, the tailrecursion > > part makes us handle intermittent stmts that were introduced by foldings > > that "clobber" our quest walking the single-use chain of stmts between > > the call and the return (and failing at all stmts that are not part > > of said chain). A simple solution is to move the stmts that are not > > part of the chain and that we can move before the call. That handles > > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c > > > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > > > > Richard. > > > > 2017-05-31 Richard Biener <rguenther@suse.de> > > > > PR middle-end/66313 > > * fold-const.c (fold_plusminus_mult_expr): If the factored > > factor may be zero use a wrapping type for the inner operation. > > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap > > and handle moved defs. > > (process_assignment): Properly guard the unary op case. Return a > > tri-state indicating that moving the stmt before the call may allow > > to continue. Pass through to_move. > > (find_tail_calls): Handle moving unrelated defs before > > the call. > > > > * c-c++-common/ubsan/pr66313.c: New testcase. > > * gcc.dg/tree-ssa/loop-15.c: Adjust. > > > > Index: gcc/fold-const.c > > =================================================================== > > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 > > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 > > *************** fold_plusminus_mult_expr (location_t loc > > *** 6916,6925 **** > > } > > same = NULL_TREE; > > > > ! if (operand_equal_p (arg01, arg11, 0)) > > ! same = arg01, alt0 = arg00, alt1 = arg10; > > ! else if (operand_equal_p (arg00, arg10, 0)) > > same = arg00, alt0 = arg01, alt1 = arg11; > > else if (operand_equal_p (arg00, arg11, 0)) > > same = arg00, alt0 = arg01, alt1 = arg10; > > else if (operand_equal_p (arg01, arg10, 0)) > > --- 6916,6926 ---- > > } > > same = NULL_TREE; > > > > ! /* Prefer factoring a common non-constant. */ > > ! if (operand_equal_p (arg00, arg10, 0)) > > same = arg00, alt0 = arg01, alt1 = arg11; > > + else if (operand_equal_p (arg01, arg11, 0)) > > + same = arg01, alt0 = arg00, alt1 = arg10; > > else if (operand_equal_p (arg00, arg11, 0)) > > same = arg00, alt0 = arg01, alt1 = arg10; > > else if (operand_equal_p (arg01, arg10, 0)) > > *************** fold_plusminus_mult_expr (location_t loc > > *** 6974,6987 **** > > } > > } > > > > ! if (same) > > return fold_build2_loc (loc, MULT_EXPR, type, > > fold_build2_loc (loc, code, type, > > fold_convert_loc (loc, type, alt0), > > fold_convert_loc (loc, type, alt1)), > > fold_convert_loc (loc, type, same)); > > > > ! return NULL_TREE; > > } > > > > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > > --- 6975,7010 ---- > > } > > } > > > > ! if (!same) > > ! return NULL_TREE; > > ! > > ! if (! INTEGRAL_TYPE_P (type) > > ! || TYPE_OVERFLOW_WRAPS (type) > > ! /* We are neither factoring zero nor minus one. */ > > ! || TREE_CODE (same) == INTEGER_CST) > > return fold_build2_loc (loc, MULT_EXPR, type, > > fold_build2_loc (loc, code, type, > > fold_convert_loc (loc, type, alt0), > > fold_convert_loc (loc, type, alt1)), > > fold_convert_loc (loc, type, same)); > > > > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise > > ! same may be minus one and thus the multiplication may overflow. Perform > > ! the operations in an unsigned type. */ > > ! tree utype = unsigned_type_for (type); > > ! tree tem = fold_build2_loc (loc, code, utype, > > ! fold_convert_loc (loc, utype, alt0), > > ! fold_convert_loc (loc, utype, alt1)); > > ! /* If the sum evaluated to a constant that is not -INF the multiplication > > ! cannot overflow. */ > > ! if (TREE_CODE (tem) == INTEGER_CST > > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) > > ! return fold_build2_loc (loc, MULT_EXPR, type, > > ! fold_convert (type, tem), same); > > ! > > ! return fold_convert_loc (loc, type, > > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, > > ! fold_convert_loc (loc, utype, same))); > > } > > > > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > > Sorry if you already know, but this part means that we can no longer > vectorise: > > int > f (int *x, int b1, int b2, int b3) > { > int foo = 0; > for (int i1 = 0; i1 < b1; ++i1) > for (int i2 = 0; i2 < b2; ++i2) > for (int i3 = 0; i3 < b3; ++i3) > foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > return foo; > } > > We now convert all the arithmetic in the [...] to unsigned int > and reassociate it so that the "- 1" is applied last. We then assume > that the overflow in this subtraction is well-defined and that the > &x[...] might not be linear for the inner loop. Can you file a PR? # i3_44 = PHI <i3_32(6), 0(4)> i3.2_7 = (unsigned int) i3_44; _9 = i3.2_7 + _34; _10 = (int) _9; _11 = (long unsigned int) _10; _12 = _11 * 4; _13 = x_30(D) + _12; _14 = *_13; ... i3_32 = i3_44 + 1; so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) It might mean that we need to avoid this re-association. Not sure how to detect worthwhile vs. not worthwhile cases during folding. At least I see no easy way to recover from it in SCEV analysis unless the number of iterations is constrained more than in your example. Ideas welcome ;) Thanks, Richard.
Richard Biener <rguenther@suse.de> writes: > On Tue, 13 Jun 2017, Richard Sandiford wrote: >> Richard Biener <rguenther@suse.de> writes: >> > So I've come back to PR66313 and found a solution to the tailrecursion >> > missed optimization when fixing the factoring folding to use an unsigned >> > type when we're not sure of overflow. >> > >> > The folding part is identical to my last try from 2015, the tailrecursion >> > part makes us handle intermittent stmts that were introduced by foldings >> > that "clobber" our quest walking the single-use chain of stmts between >> > the call and the return (and failing at all stmts that are not part >> > of said chain). A simple solution is to move the stmts that are not >> > part of the chain and that we can move before the call. That handles >> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c >> > >> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. >> > >> > Richard. >> > >> > 2017-05-31 Richard Biener <rguenther@suse.de> >> > >> > PR middle-end/66313 >> > * fold-const.c (fold_plusminus_mult_expr): If the factored >> > factor may be zero use a wrapping type for the inner operation. >> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap >> > and handle moved defs. >> > (process_assignment): Properly guard the unary op case. Return a >> > tri-state indicating that moving the stmt before the call may allow >> > to continue. Pass through to_move. >> > (find_tail_calls): Handle moving unrelated defs before >> > the call. >> > >> > * c-c++-common/ubsan/pr66313.c: New testcase. >> > * gcc.dg/tree-ssa/loop-15.c: Adjust. >> > >> > Index: gcc/fold-const.c >> > =================================================================== >> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 >> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 >> > *************** fold_plusminus_mult_expr (location_t loc >> > *** 6916,6925 **** >> > } >> > same = NULL_TREE; >> > >> > ! if (operand_equal_p (arg01, arg11, 0)) >> > ! same = arg01, alt0 = arg00, alt1 = arg10; >> > ! else if (operand_equal_p (arg00, arg10, 0)) >> > same = arg00, alt0 = arg01, alt1 = arg11; >> > else if (operand_equal_p (arg00, arg11, 0)) >> > same = arg00, alt0 = arg01, alt1 = arg10; >> > else if (operand_equal_p (arg01, arg10, 0)) >> > --- 6916,6926 ---- >> > } >> > same = NULL_TREE; >> > >> > ! /* Prefer factoring a common non-constant. */ >> > ! if (operand_equal_p (arg00, arg10, 0)) >> > same = arg00, alt0 = arg01, alt1 = arg11; >> > + else if (operand_equal_p (arg01, arg11, 0)) >> > + same = arg01, alt0 = arg00, alt1 = arg10; >> > else if (operand_equal_p (arg00, arg11, 0)) >> > same = arg00, alt0 = arg01, alt1 = arg10; >> > else if (operand_equal_p (arg01, arg10, 0)) >> > *************** fold_plusminus_mult_expr (location_t loc >> > *** 6974,6987 **** >> > } >> > } >> > >> > ! if (same) >> > return fold_build2_loc (loc, MULT_EXPR, type, >> > fold_build2_loc (loc, code, type, >> > fold_convert_loc (loc, type, alt0), >> > fold_convert_loc (loc, type, alt1)), >> > fold_convert_loc (loc, type, same)); >> > >> > ! return NULL_TREE; >> > } >> > >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >> > --- 6975,7010 ---- >> > } >> > } >> > >> > ! if (!same) >> > ! return NULL_TREE; >> > ! >> > ! if (! INTEGRAL_TYPE_P (type) >> > ! || TYPE_OVERFLOW_WRAPS (type) >> > ! /* We are neither factoring zero nor minus one. */ >> > ! || TREE_CODE (same) == INTEGER_CST) >> > return fold_build2_loc (loc, MULT_EXPR, type, >> > fold_build2_loc (loc, code, type, >> > fold_convert_loc (loc, type, alt0), >> > fold_convert_loc (loc, type, alt1)), >> > fold_convert_loc (loc, type, same)); >> > >> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise >> > ! same may be minus one and thus the multiplication may overflow. Perform >> > ! the operations in an unsigned type. */ >> > ! tree utype = unsigned_type_for (type); >> > ! tree tem = fold_build2_loc (loc, code, utype, >> > ! fold_convert_loc (loc, utype, alt0), >> > ! fold_convert_loc (loc, utype, alt1)); >> > ! /* If the sum evaluated to a constant that is not -INF the multiplication >> > ! cannot overflow. */ >> > ! if (TREE_CODE (tem) == INTEGER_CST >> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) >> > ! return fold_build2_loc (loc, MULT_EXPR, type, >> > ! fold_convert (type, tem), same); >> > ! >> > ! return fold_convert_loc (loc, type, >> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, >> > ! fold_convert_loc (loc, utype, same))); >> > } >> > >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >> >> Sorry if you already know, but this part means that we can no longer >> vectorise: >> >> int >> f (int *x, int b1, int b2, int b3) >> { >> int foo = 0; >> for (int i1 = 0; i1 < b1; ++i1) >> for (int i2 = 0; i2 < b2; ++i2) >> for (int i3 = 0; i3 < b3; ++i3) >> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; >> return foo; >> } >> >> We now convert all the arithmetic in the [...] to unsigned int >> and reassociate it so that the "- 1" is applied last. We then assume >> that the overflow in this subtraction is well-defined and that the >> &x[...] might not be linear for the inner loop. > > Can you file a PR? OK, filed as PR81082. > # i3_44 = PHI <i3_32(6), 0(4)> > i3.2_7 = (unsigned int) i3_44; > _9 = i3.2_7 + _34; > _10 = (int) _9; > _11 = (long unsigned int) _10; > _12 = _11 * 4; > _13 = x_30(D) + _12; > _14 = *_13; > ... > i3_32 = i3_44 + 1; > > so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) > > It might mean that we need to avoid this re-association. Not sure how > to detect worthwhile vs. not worthwhile cases during folding. At least > I see no easy way to recover from it in SCEV analysis unless the > number of iterations is constrained more than in your example. Yeah, in the example that this was reduced from, not reassociating is far more preferable to changing the types. But like you say, I've no idea how we'd know that at this stage. Thanks, Richard > Ideas welcome ;) > > Thanks, > Richard.
On Tue, Jun 13, 2017 at 12:23 PM, Richard Sandiford <richard.sandiford@linaro.org> wrote: > Richard Biener <rguenther@suse.de> writes: >> On Tue, 13 Jun 2017, Richard Sandiford wrote: >>> Richard Biener <rguenther@suse.de> writes: >>> > So I've come back to PR66313 and found a solution to the tailrecursion >>> > missed optimization when fixing the factoring folding to use an unsigned >>> > type when we're not sure of overflow. >>> > >>> > The folding part is identical to my last try from 2015, the tailrecursion >>> > part makes us handle intermittent stmts that were introduced by foldings >>> > that "clobber" our quest walking the single-use chain of stmts between >>> > the call and the return (and failing at all stmts that are not part >>> > of said chain). A simple solution is to move the stmts that are not >>> > part of the chain and that we can move before the call. That handles >>> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c >>> > >>> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. >>> > >>> > Richard. >>> > >>> > 2017-05-31 Richard Biener <rguenther@suse.de> >>> > >>> > PR middle-end/66313 >>> > * fold-const.c (fold_plusminus_mult_expr): If the factored >>> > factor may be zero use a wrapping type for the inner operation. >>> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap >>> > and handle moved defs. >>> > (process_assignment): Properly guard the unary op case. Return a >>> > tri-state indicating that moving the stmt before the call may allow >>> > to continue. Pass through to_move. >>> > (find_tail_calls): Handle moving unrelated defs before >>> > the call. >>> > >>> > * c-c++-common/ubsan/pr66313.c: New testcase. >>> > * gcc.dg/tree-ssa/loop-15.c: Adjust. >>> > >>> > Index: gcc/fold-const.c >>> > =================================================================== >>> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 >>> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 >>> > *************** fold_plusminus_mult_expr (location_t loc >>> > *** 6916,6925 **** >>> > } >>> > same = NULL_TREE; >>> > >>> > ! if (operand_equal_p (arg01, arg11, 0)) >>> > ! same = arg01, alt0 = arg00, alt1 = arg10; >>> > ! else if (operand_equal_p (arg00, arg10, 0)) >>> > same = arg00, alt0 = arg01, alt1 = arg11; >>> > else if (operand_equal_p (arg00, arg11, 0)) >>> > same = arg00, alt0 = arg01, alt1 = arg10; >>> > else if (operand_equal_p (arg01, arg10, 0)) >>> > --- 6916,6926 ---- >>> > } >>> > same = NULL_TREE; >>> > >>> > ! /* Prefer factoring a common non-constant. */ >>> > ! if (operand_equal_p (arg00, arg10, 0)) >>> > same = arg00, alt0 = arg01, alt1 = arg11; >>> > + else if (operand_equal_p (arg01, arg11, 0)) >>> > + same = arg01, alt0 = arg00, alt1 = arg10; >>> > else if (operand_equal_p (arg00, arg11, 0)) >>> > same = arg00, alt0 = arg01, alt1 = arg10; >>> > else if (operand_equal_p (arg01, arg10, 0)) >>> > *************** fold_plusminus_mult_expr (location_t loc >>> > *** 6974,6987 **** >>> > } >>> > } >>> > >>> > ! if (same) >>> > return fold_build2_loc (loc, MULT_EXPR, type, >>> > fold_build2_loc (loc, code, type, >>> > fold_convert_loc (loc, type, alt0), >>> > fold_convert_loc (loc, type, alt1)), >>> > fold_convert_loc (loc, type, same)); >>> > >>> > ! return NULL_TREE; >>> > } >>> > >>> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >>> > --- 6975,7010 ---- >>> > } >>> > } >>> > >>> > ! if (!same) >>> > ! return NULL_TREE; >>> > ! >>> > ! if (! INTEGRAL_TYPE_P (type) >>> > ! || TYPE_OVERFLOW_WRAPS (type) >>> > ! /* We are neither factoring zero nor minus one. */ >>> > ! || TREE_CODE (same) == INTEGER_CST) >>> > return fold_build2_loc (loc, MULT_EXPR, type, >>> > fold_build2_loc (loc, code, type, >>> > fold_convert_loc (loc, type, alt0), >>> > fold_convert_loc (loc, type, alt1)), >>> > fold_convert_loc (loc, type, same)); >>> > >>> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise >>> > ! same may be minus one and thus the multiplication may overflow. Perform >>> > ! the operations in an unsigned type. */ >>> > ! tree utype = unsigned_type_for (type); >>> > ! tree tem = fold_build2_loc (loc, code, utype, >>> > ! fold_convert_loc (loc, utype, alt0), >>> > ! fold_convert_loc (loc, utype, alt1)); >>> > ! /* If the sum evaluated to a constant that is not -INF the multiplication >>> > ! cannot overflow. */ >>> > ! if (TREE_CODE (tem) == INTEGER_CST >>> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) >>> > ! return fold_build2_loc (loc, MULT_EXPR, type, >>> > ! fold_convert (type, tem), same); >>> > ! >>> > ! return fold_convert_loc (loc, type, >>> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, >>> > ! fold_convert_loc (loc, utype, same))); >>> > } >>> > >>> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >>> >>> Sorry if you already know, but this part means that we can no longer >>> vectorise: >>> >>> int >>> f (int *x, int b1, int b2, int b3) >>> { >>> int foo = 0; >>> for (int i1 = 0; i1 < b1; ++i1) >>> for (int i2 = 0; i2 < b2; ++i2) >>> for (int i3 = 0; i3 < b3; ++i3) >>> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; >>> return foo; >>> } >>> >>> We now convert all the arithmetic in the [...] to unsigned int >>> and reassociate it so that the "- 1" is applied last. We then assume >>> that the overflow in this subtraction is well-defined and that the >>> &x[...] might not be linear for the inner loop. >> >> Can you file a PR? > > OK, filed as PR81082. > >> # i3_44 = PHI <i3_32(6), 0(4)> >> i3.2_7 = (unsigned int) i3_44; >> _9 = i3.2_7 + _34; >> _10 = (int) _9; >> _11 = (long unsigned int) _10; >> _12 = _11 * 4; >> _13 = x_30(D) + _12; >> _14 = *_13; >> ... >> i3_32 = i3_44 + 1; >> >> so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) >> >> It might mean that we need to avoid this re-association. Not sure how >> to detect worthwhile vs. not worthwhile cases during folding. At least >> I see no easy way to recover from it in SCEV analysis unless the >> number of iterations is constrained more than in your example. > > Yeah, in the example that this was reduced from, not reassociating > is far more preferable to changing the types. But like you say, > I've no idea how we'd know that at this stage. Looks like it has become a general problem. IIRC, loop-im also hoists signed computation out of loop into unsigned computations. Thanks, bin > > Thanks, > Richard > >> Ideas welcome ;) >> >> Thanks, >> Richard.
On Tue, 13 Jun 2017, Richard Sandiford wrote: > Richard Biener <rguenther@suse.de> writes: > > On Tue, 13 Jun 2017, Richard Sandiford wrote: > >> Richard Biener <rguenther@suse.de> writes: > >> > So I've come back to PR66313 and found a solution to the tailrecursion > >> > missed optimization when fixing the factoring folding to use an unsigned > >> > type when we're not sure of overflow. > >> > > >> > The folding part is identical to my last try from 2015, the tailrecursion > >> > part makes us handle intermittent stmts that were introduced by foldings > >> > that "clobber" our quest walking the single-use chain of stmts between > >> > the call and the return (and failing at all stmts that are not part > >> > of said chain). A simple solution is to move the stmts that are not > >> > part of the chain and that we can move before the call. That handles > >> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c > >> > > >> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > >> > > >> > Richard. > >> > > >> > 2017-05-31 Richard Biener <rguenther@suse.de> > >> > > >> > PR middle-end/66313 > >> > * fold-const.c (fold_plusminus_mult_expr): If the factored > >> > factor may be zero use a wrapping type for the inner operation. > >> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap > >> > and handle moved defs. > >> > (process_assignment): Properly guard the unary op case. Return a > >> > tri-state indicating that moving the stmt before the call may allow > >> > to continue. Pass through to_move. > >> > (find_tail_calls): Handle moving unrelated defs before > >> > the call. > >> > > >> > * c-c++-common/ubsan/pr66313.c: New testcase. > >> > * gcc.dg/tree-ssa/loop-15.c: Adjust. > >> > > >> > Index: gcc/fold-const.c > >> > =================================================================== > >> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 > >> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 > >> > *************** fold_plusminus_mult_expr (location_t loc > >> > *** 6916,6925 **** > >> > } > >> > same = NULL_TREE; > >> > > >> > ! if (operand_equal_p (arg01, arg11, 0)) > >> > ! same = arg01, alt0 = arg00, alt1 = arg10; > >> > ! else if (operand_equal_p (arg00, arg10, 0)) > >> > same = arg00, alt0 = arg01, alt1 = arg11; > >> > else if (operand_equal_p (arg00, arg11, 0)) > >> > same = arg00, alt0 = arg01, alt1 = arg10; > >> > else if (operand_equal_p (arg01, arg10, 0)) > >> > --- 6916,6926 ---- > >> > } > >> > same = NULL_TREE; > >> > > >> > ! /* Prefer factoring a common non-constant. */ > >> > ! if (operand_equal_p (arg00, arg10, 0)) > >> > same = arg00, alt0 = arg01, alt1 = arg11; > >> > + else if (operand_equal_p (arg01, arg11, 0)) > >> > + same = arg01, alt0 = arg00, alt1 = arg10; > >> > else if (operand_equal_p (arg00, arg11, 0)) > >> > same = arg00, alt0 = arg01, alt1 = arg10; > >> > else if (operand_equal_p (arg01, arg10, 0)) > >> > *************** fold_plusminus_mult_expr (location_t loc > >> > *** 6974,6987 **** > >> > } > >> > } > >> > > >> > ! if (same) > >> > return fold_build2_loc (loc, MULT_EXPR, type, > >> > fold_build2_loc (loc, code, type, > >> > fold_convert_loc (loc, type, alt0), > >> > fold_convert_loc (loc, type, alt1)), > >> > fold_convert_loc (loc, type, same)); > >> > > >> > ! return NULL_TREE; > >> > } > >> > > >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >> > --- 6975,7010 ---- > >> > } > >> > } > >> > > >> > ! if (!same) > >> > ! return NULL_TREE; > >> > ! > >> > ! if (! INTEGRAL_TYPE_P (type) > >> > ! || TYPE_OVERFLOW_WRAPS (type) > >> > ! /* We are neither factoring zero nor minus one. */ > >> > ! || TREE_CODE (same) == INTEGER_CST) > >> > return fold_build2_loc (loc, MULT_EXPR, type, > >> > fold_build2_loc (loc, code, type, > >> > fold_convert_loc (loc, type, alt0), > >> > fold_convert_loc (loc, type, alt1)), > >> > fold_convert_loc (loc, type, same)); > >> > > >> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise > >> > ! same may be minus one and thus the multiplication may overflow. Perform > >> > ! the operations in an unsigned type. */ > >> > ! tree utype = unsigned_type_for (type); > >> > ! tree tem = fold_build2_loc (loc, code, utype, > >> > ! fold_convert_loc (loc, utype, alt0), > >> > ! fold_convert_loc (loc, utype, alt1)); > >> > ! /* If the sum evaluated to a constant that is not -INF the multiplication > >> > ! cannot overflow. */ > >> > ! if (TREE_CODE (tem) == INTEGER_CST > >> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) > >> > ! return fold_build2_loc (loc, MULT_EXPR, type, > >> > ! fold_convert (type, tem), same); > >> > ! > >> > ! return fold_convert_loc (loc, type, > >> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, > >> > ! fold_convert_loc (loc, utype, same))); > >> > } > >> > > >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >> > >> Sorry if you already know, but this part means that we can no longer > >> vectorise: > >> > >> int > >> f (int *x, int b1, int b2, int b3) > >> { > >> int foo = 0; > >> for (int i1 = 0; i1 < b1; ++i1) > >> for (int i2 = 0; i2 < b2; ++i2) > >> for (int i3 = 0; i3 < b3; ++i3) > >> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > >> return foo; > >> } > >> > >> We now convert all the arithmetic in the [...] to unsigned int > >> and reassociate it so that the "- 1" is applied last. We then assume > >> that the overflow in this subtraction is well-defined and that the > >> &x[...] might not be linear for the inner loop. > > > > Can you file a PR? > > OK, filed as PR81082. > > > # i3_44 = PHI <i3_32(6), 0(4)> > > i3.2_7 = (unsigned int) i3_44; > > _9 = i3.2_7 + _34; > > _10 = (int) _9; > > _11 = (long unsigned int) _10; > > _12 = _11 * 4; > > _13 = x_30(D) + _12; > > _14 = *_13; > > ... > > i3_32 = i3_44 + 1; > > > > so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) > > > > It might mean that we need to avoid this re-association. Not sure how > > to detect worthwhile vs. not worthwhile cases during folding. At least > > I see no easy way to recover from it in SCEV analysis unless the > > number of iterations is constrained more than in your example. > > Yeah, in the example that this was reduced from, not reassociating > is far more preferable to changing the types. But like you say, > I've no idea how we'd know that at this stage. In the past I played with not obfuscating things during FE lowering so early, namely not expose the * 4 but keep the original types of array / pointer indexes. On the original mem-ref branch I had a IDX_EXPR that allowed to chain those (a widen-multiply-accumulate op). That said, the context where this association is not interesting is address context and the multiplication to the byte offset. Of course in your case there's also b3 factored out which looks like a generally interesting transform (but eventually harmful in address context again). From the FE we get *(x + (sizetype) ((long unsigned int) (((i1 * b2) * b3 + i2 * b3) + (i3 + -1)) * 4)) and SCEV uses the fact that the mults/adds have undefined behavior on overflow. The same missed optimization occurs if you make all those vars unsigned (with or without the folding in question): #define U unsigned int f (int *x, U int b1, U int b2, U int b3) { int foo = 0; for (U int i1 = 0; i1 < b1; ++i1) for (U int i2 = 0; i2 < b2; ++i2) for (U int i3 = 0; i3 < b3; ++i3) foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; return foo; } It works again for unsigned long as there can be no valid object where the computation could wrap around. Richard.
On Tue, 13 Jun 2017, Bin.Cheng wrote: > On Tue, Jun 13, 2017 at 12:23 PM, Richard Sandiford > <richard.sandiford@linaro.org> wrote: > > Richard Biener <rguenther@suse.de> writes: > >> On Tue, 13 Jun 2017, Richard Sandiford wrote: > >>> Richard Biener <rguenther@suse.de> writes: > >>> > So I've come back to PR66313 and found a solution to the tailrecursion > >>> > missed optimization when fixing the factoring folding to use an unsigned > >>> > type when we're not sure of overflow. > >>> > > >>> > The folding part is identical to my last try from 2015, the tailrecursion > >>> > part makes us handle intermittent stmts that were introduced by foldings > >>> > that "clobber" our quest walking the single-use chain of stmts between > >>> > the call and the return (and failing at all stmts that are not part > >>> > of said chain). A simple solution is to move the stmts that are not > >>> > part of the chain and that we can move before the call. That handles > >>> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c > >>> > > >>> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > >>> > > >>> > Richard. > >>> > > >>> > 2017-05-31 Richard Biener <rguenther@suse.de> > >>> > > >>> > PR middle-end/66313 > >>> > * fold-const.c (fold_plusminus_mult_expr): If the factored > >>> > factor may be zero use a wrapping type for the inner operation. > >>> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap > >>> > and handle moved defs. > >>> > (process_assignment): Properly guard the unary op case. Return a > >>> > tri-state indicating that moving the stmt before the call may allow > >>> > to continue. Pass through to_move. > >>> > (find_tail_calls): Handle moving unrelated defs before > >>> > the call. > >>> > > >>> > * c-c++-common/ubsan/pr66313.c: New testcase. > >>> > * gcc.dg/tree-ssa/loop-15.c: Adjust. > >>> > > >>> > Index: gcc/fold-const.c > >>> > =================================================================== > >>> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 > >>> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 > >>> > *************** fold_plusminus_mult_expr (location_t loc > >>> > *** 6916,6925 **** > >>> > } > >>> > same = NULL_TREE; > >>> > > >>> > ! if (operand_equal_p (arg01, arg11, 0)) > >>> > ! same = arg01, alt0 = arg00, alt1 = arg10; > >>> > ! else if (operand_equal_p (arg00, arg10, 0)) > >>> > same = arg00, alt0 = arg01, alt1 = arg11; > >>> > else if (operand_equal_p (arg00, arg11, 0)) > >>> > same = arg00, alt0 = arg01, alt1 = arg10; > >>> > else if (operand_equal_p (arg01, arg10, 0)) > >>> > --- 6916,6926 ---- > >>> > } > >>> > same = NULL_TREE; > >>> > > >>> > ! /* Prefer factoring a common non-constant. */ > >>> > ! if (operand_equal_p (arg00, arg10, 0)) > >>> > same = arg00, alt0 = arg01, alt1 = arg11; > >>> > + else if (operand_equal_p (arg01, arg11, 0)) > >>> > + same = arg01, alt0 = arg00, alt1 = arg10; > >>> > else if (operand_equal_p (arg00, arg11, 0)) > >>> > same = arg00, alt0 = arg01, alt1 = arg10; > >>> > else if (operand_equal_p (arg01, arg10, 0)) > >>> > *************** fold_plusminus_mult_expr (location_t loc > >>> > *** 6974,6987 **** > >>> > } > >>> > } > >>> > > >>> > ! if (same) > >>> > return fold_build2_loc (loc, MULT_EXPR, type, > >>> > fold_build2_loc (loc, code, type, > >>> > fold_convert_loc (loc, type, alt0), > >>> > fold_convert_loc (loc, type, alt1)), > >>> > fold_convert_loc (loc, type, same)); > >>> > > >>> > ! return NULL_TREE; > >>> > } > >>> > > >>> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >>> > --- 6975,7010 ---- > >>> > } > >>> > } > >>> > > >>> > ! if (!same) > >>> > ! return NULL_TREE; > >>> > ! > >>> > ! if (! INTEGRAL_TYPE_P (type) > >>> > ! || TYPE_OVERFLOW_WRAPS (type) > >>> > ! /* We are neither factoring zero nor minus one. */ > >>> > ! || TREE_CODE (same) == INTEGER_CST) > >>> > return fold_build2_loc (loc, MULT_EXPR, type, > >>> > fold_build2_loc (loc, code, type, > >>> > fold_convert_loc (loc, type, alt0), > >>> > fold_convert_loc (loc, type, alt1)), > >>> > fold_convert_loc (loc, type, same)); > >>> > > >>> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise > >>> > ! same may be minus one and thus the multiplication may overflow. Perform > >>> > ! the operations in an unsigned type. */ > >>> > ! tree utype = unsigned_type_for (type); > >>> > ! tree tem = fold_build2_loc (loc, code, utype, > >>> > ! fold_convert_loc (loc, utype, alt0), > >>> > ! fold_convert_loc (loc, utype, alt1)); > >>> > ! /* If the sum evaluated to a constant that is not -INF the multiplication > >>> > ! cannot overflow. */ > >>> > ! if (TREE_CODE (tem) == INTEGER_CST > >>> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) > >>> > ! return fold_build2_loc (loc, MULT_EXPR, type, > >>> > ! fold_convert (type, tem), same); > >>> > ! > >>> > ! return fold_convert_loc (loc, type, > >>> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, > >>> > ! fold_convert_loc (loc, utype, same))); > >>> > } > >>> > > >>> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >>> > >>> Sorry if you already know, but this part means that we can no longer > >>> vectorise: > >>> > >>> int > >>> f (int *x, int b1, int b2, int b3) > >>> { > >>> int foo = 0; > >>> for (int i1 = 0; i1 < b1; ++i1) > >>> for (int i2 = 0; i2 < b2; ++i2) > >>> for (int i3 = 0; i3 < b3; ++i3) > >>> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > >>> return foo; > >>> } > >>> > >>> We now convert all the arithmetic in the [...] to unsigned int > >>> and reassociate it so that the "- 1" is applied last. We then assume > >>> that the overflow in this subtraction is well-defined and that the > >>> &x[...] might not be linear for the inner loop. > >> > >> Can you file a PR? > > > > OK, filed as PR81082. > > > >> # i3_44 = PHI <i3_32(6), 0(4)> > >> i3.2_7 = (unsigned int) i3_44; > >> _9 = i3.2_7 + _34; > >> _10 = (int) _9; > >> _11 = (long unsigned int) _10; > >> _12 = _11 * 4; > >> _13 = x_30(D) + _12; > >> _14 = *_13; > >> ... > >> i3_32 = i3_44 + 1; > >> > >> so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) > >> > >> It might mean that we need to avoid this re-association. Not sure how > >> to detect worthwhile vs. not worthwhile cases during folding. At least > >> I see no easy way to recover from it in SCEV analysis unless the > >> number of iterations is constrained more than in your example. > > > > Yeah, in the example that this was reduced from, not reassociating > > is far more preferable to changing the types. But like you say, > > I've no idea how we'd know that at this stage. > Looks like it has become a general problem. IIRC, loop-im also hoists > signed computation out of loop into unsigned computations. Yes, it's the trade off of having signed ops invoke undefined behavior and not spuriously introducing this ourselves by 1) not doing anything 2) being careful and re-writing to arith not involving undefined behavior. Richard. > Thanks, > bin > > > > Thanks, > > Richard > > > >> Ideas welcome ;) > >> > >> Thanks, > >> Richard. > >
On Tue, Jun 13, 2017 at 12:48 PM, Richard Biener <rguenther@suse.de> wrote: > On Tue, 13 Jun 2017, Richard Sandiford wrote: > >> Richard Biener <rguenther@suse.de> writes: >> > On Tue, 13 Jun 2017, Richard Sandiford wrote: >> >> Richard Biener <rguenther@suse.de> writes: >> >> > So I've come back to PR66313 and found a solution to the tailrecursion >> >> > missed optimization when fixing the factoring folding to use an unsigned >> >> > type when we're not sure of overflow. >> >> > >> >> > The folding part is identical to my last try from 2015, the tailrecursion >> >> > part makes us handle intermittent stmts that were introduced by foldings >> >> > that "clobber" our quest walking the single-use chain of stmts between >> >> > the call and the return (and failing at all stmts that are not part >> >> > of said chain). A simple solution is to move the stmts that are not >> >> > part of the chain and that we can move before the call. That handles >> >> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c >> >> > >> >> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. >> >> > >> >> > Richard. >> >> > >> >> > 2017-05-31 Richard Biener <rguenther@suse.de> >> >> > >> >> > PR middle-end/66313 >> >> > * fold-const.c (fold_plusminus_mult_expr): If the factored >> >> > factor may be zero use a wrapping type for the inner operation. >> >> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap >> >> > and handle moved defs. >> >> > (process_assignment): Properly guard the unary op case. Return a >> >> > tri-state indicating that moving the stmt before the call may allow >> >> > to continue. Pass through to_move. >> >> > (find_tail_calls): Handle moving unrelated defs before >> >> > the call. >> >> > >> >> > * c-c++-common/ubsan/pr66313.c: New testcase. >> >> > * gcc.dg/tree-ssa/loop-15.c: Adjust. >> >> > >> >> > Index: gcc/fold-const.c >> >> > =================================================================== >> >> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 >> >> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 >> >> > *************** fold_plusminus_mult_expr (location_t loc >> >> > *** 6916,6925 **** >> >> > } >> >> > same = NULL_TREE; >> >> > >> >> > ! if (operand_equal_p (arg01, arg11, 0)) >> >> > ! same = arg01, alt0 = arg00, alt1 = arg10; >> >> > ! else if (operand_equal_p (arg00, arg10, 0)) >> >> > same = arg00, alt0 = arg01, alt1 = arg11; >> >> > else if (operand_equal_p (arg00, arg11, 0)) >> >> > same = arg00, alt0 = arg01, alt1 = arg10; >> >> > else if (operand_equal_p (arg01, arg10, 0)) >> >> > --- 6916,6926 ---- >> >> > } >> >> > same = NULL_TREE; >> >> > >> >> > ! /* Prefer factoring a common non-constant. */ >> >> > ! if (operand_equal_p (arg00, arg10, 0)) >> >> > same = arg00, alt0 = arg01, alt1 = arg11; >> >> > + else if (operand_equal_p (arg01, arg11, 0)) >> >> > + same = arg01, alt0 = arg00, alt1 = arg10; >> >> > else if (operand_equal_p (arg00, arg11, 0)) >> >> > same = arg00, alt0 = arg01, alt1 = arg10; >> >> > else if (operand_equal_p (arg01, arg10, 0)) >> >> > *************** fold_plusminus_mult_expr (location_t loc >> >> > *** 6974,6987 **** >> >> > } >> >> > } >> >> > >> >> > ! if (same) >> >> > return fold_build2_loc (loc, MULT_EXPR, type, >> >> > fold_build2_loc (loc, code, type, >> >> > fold_convert_loc (loc, type, alt0), >> >> > fold_convert_loc (loc, type, alt1)), >> >> > fold_convert_loc (loc, type, same)); >> >> > >> >> > ! return NULL_TREE; >> >> > } >> >> > >> >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >> >> > --- 6975,7010 ---- >> >> > } >> >> > } >> >> > >> >> > ! if (!same) >> >> > ! return NULL_TREE; >> >> > ! >> >> > ! if (! INTEGRAL_TYPE_P (type) >> >> > ! || TYPE_OVERFLOW_WRAPS (type) >> >> > ! /* We are neither factoring zero nor minus one. */ >> >> > ! || TREE_CODE (same) == INTEGER_CST) >> >> > return fold_build2_loc (loc, MULT_EXPR, type, >> >> > fold_build2_loc (loc, code, type, >> >> > fold_convert_loc (loc, type, alt0), >> >> > fold_convert_loc (loc, type, alt1)), >> >> > fold_convert_loc (loc, type, same)); >> >> > >> >> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise >> >> > ! same may be minus one and thus the multiplication may overflow. Perform >> >> > ! the operations in an unsigned type. */ >> >> > ! tree utype = unsigned_type_for (type); >> >> > ! tree tem = fold_build2_loc (loc, code, utype, >> >> > ! fold_convert_loc (loc, utype, alt0), >> >> > ! fold_convert_loc (loc, utype, alt1)); >> >> > ! /* If the sum evaluated to a constant that is not -INF the multiplication >> >> > ! cannot overflow. */ >> >> > ! if (TREE_CODE (tem) == INTEGER_CST >> >> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) >> >> > ! return fold_build2_loc (loc, MULT_EXPR, type, >> >> > ! fold_convert (type, tem), same); >> >> > ! >> >> > ! return fold_convert_loc (loc, type, >> >> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, >> >> > ! fold_convert_loc (loc, utype, same))); >> >> > } >> >> > >> >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST >> >> >> >> Sorry if you already know, but this part means that we can no longer >> >> vectorise: >> >> >> >> int >> >> f (int *x, int b1, int b2, int b3) >> >> { >> >> int foo = 0; >> >> for (int i1 = 0; i1 < b1; ++i1) >> >> for (int i2 = 0; i2 < b2; ++i2) >> >> for (int i3 = 0; i3 < b3; ++i3) >> >> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; >> >> return foo; >> >> } >> >> >> >> We now convert all the arithmetic in the [...] to unsigned int >> >> and reassociate it so that the "- 1" is applied last. We then assume >> >> that the overflow in this subtraction is well-defined and that the >> >> &x[...] might not be linear for the inner loop. >> > >> > Can you file a PR? >> >> OK, filed as PR81082. >> >> > # i3_44 = PHI <i3_32(6), 0(4)> >> > i3.2_7 = (unsigned int) i3_44; >> > _9 = i3.2_7 + _34; >> > _10 = (int) _9; >> > _11 = (long unsigned int) _10; >> > _12 = _11 * 4; >> > _13 = x_30(D) + _12; >> > _14 = *_13; >> > ... >> > i3_32 = i3_44 + 1; >> > >> > so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) >> > >> > It might mean that we need to avoid this re-association. Not sure how >> > to detect worthwhile vs. not worthwhile cases during folding. At least >> > I see no easy way to recover from it in SCEV analysis unless the >> > number of iterations is constrained more than in your example. >> >> Yeah, in the example that this was reduced from, not reassociating >> is far more preferable to changing the types. But like you say, >> I've no idea how we'd know that at this stage. > > In the past I played with not obfuscating things during FE lowering > so early, namely not expose the * 4 but keep the original types of > array / pointer indexes. On the original mem-ref branch I had > a IDX_EXPR that allowed to chain those (a widen-multiply-accumulate > op). > > That said, the context where this association is not interesting is > address context and the multiplication to the byte offset. > > Of course in your case there's also b3 factored out which looks > like a generally interesting transform (but eventually harmful in address > context again). > > From the FE we get > > *(x + (sizetype) ((long unsigned int) (((i1 * b2) * b3 + i2 * b3) + (i3 + > -1)) * 4)) > > and SCEV uses the fact that the mults/adds have undefined behavior > on overflow. The same missed optimization occurs if you make all those > vars unsigned (with or without the folding in question): But with unsigned type it's not missed optimization because computation could overflow and result in non-scev address. In this case the loop needs to be versioned under overflow/non-overflow to be parallelized/vectorized, right? Or split for cases we can easily infer boundary condition of overflow. Thanks, bin > > #define U unsigned > int > f (int *x, U int b1, U int b2, U int b3) > { > int foo = 0; > for (U int i1 = 0; i1 < b1; ++i1) > for (U int i2 = 0; i2 < b2; ++i2) > for (U int i3 = 0; i3 < b3; ++i3) > foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > return foo; > } > > It works again for unsigned long as there can be no valid object > where the computation could wrap around. > > Richard. > > -- > Richard Biener <rguenther@suse.de> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
On Tue, 13 Jun 2017, Bin.Cheng wrote: > On Tue, Jun 13, 2017 at 12:48 PM, Richard Biener <rguenther@suse.de> wrote: > > On Tue, 13 Jun 2017, Richard Sandiford wrote: > > > >> Richard Biener <rguenther@suse.de> writes: > >> > On Tue, 13 Jun 2017, Richard Sandiford wrote: > >> >> Richard Biener <rguenther@suse.de> writes: > >> >> > So I've come back to PR66313 and found a solution to the tailrecursion > >> >> > missed optimization when fixing the factoring folding to use an unsigned > >> >> > type when we're not sure of overflow. > >> >> > > >> >> > The folding part is identical to my last try from 2015, the tailrecursion > >> >> > part makes us handle intermittent stmts that were introduced by foldings > >> >> > that "clobber" our quest walking the single-use chain of stmts between > >> >> > the call and the return (and failing at all stmts that are not part > >> >> > of said chain). A simple solution is to move the stmts that are not > >> >> > part of the chain and that we can move before the call. That handles > >> >> > the leaf conversions that now appear for tree-ssa/tailrecursion-6.c > >> >> > > >> >> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > >> >> > > >> >> > Richard. > >> >> > > >> >> > 2017-05-31 Richard Biener <rguenther@suse.de> > >> >> > > >> >> > PR middle-end/66313 > >> >> > * fold-const.c (fold_plusminus_mult_expr): If the factored > >> >> > factor may be zero use a wrapping type for the inner operation. > >> >> > * tree-tailcall.c (independent_of_stmt_p): Pass in to_move bitmap > >> >> > and handle moved defs. > >> >> > (process_assignment): Properly guard the unary op case. Return a > >> >> > tri-state indicating that moving the stmt before the call may allow > >> >> > to continue. Pass through to_move. > >> >> > (find_tail_calls): Handle moving unrelated defs before > >> >> > the call. > >> >> > > >> >> > * c-c++-common/ubsan/pr66313.c: New testcase. > >> >> > * gcc.dg/tree-ssa/loop-15.c: Adjust. > >> >> > > >> >> > Index: gcc/fold-const.c > >> >> > =================================================================== > >> >> > *** gcc/fold-const.c.orig 2015-10-29 12:32:33.302782318 +0100 > >> >> > --- gcc/fold-const.c 2015-10-29 14:08:39.936497739 +0100 > >> >> > *************** fold_plusminus_mult_expr (location_t loc > >> >> > *** 6916,6925 **** > >> >> > } > >> >> > same = NULL_TREE; > >> >> > > >> >> > ! if (operand_equal_p (arg01, arg11, 0)) > >> >> > ! same = arg01, alt0 = arg00, alt1 = arg10; > >> >> > ! else if (operand_equal_p (arg00, arg10, 0)) > >> >> > same = arg00, alt0 = arg01, alt1 = arg11; > >> >> > else if (operand_equal_p (arg00, arg11, 0)) > >> >> > same = arg00, alt0 = arg01, alt1 = arg10; > >> >> > else if (operand_equal_p (arg01, arg10, 0)) > >> >> > --- 6916,6926 ---- > >> >> > } > >> >> > same = NULL_TREE; > >> >> > > >> >> > ! /* Prefer factoring a common non-constant. */ > >> >> > ! if (operand_equal_p (arg00, arg10, 0)) > >> >> > same = arg00, alt0 = arg01, alt1 = arg11; > >> >> > + else if (operand_equal_p (arg01, arg11, 0)) > >> >> > + same = arg01, alt0 = arg00, alt1 = arg10; > >> >> > else if (operand_equal_p (arg00, arg11, 0)) > >> >> > same = arg00, alt0 = arg01, alt1 = arg10; > >> >> > else if (operand_equal_p (arg01, arg10, 0)) > >> >> > *************** fold_plusminus_mult_expr (location_t loc > >> >> > *** 6974,6987 **** > >> >> > } > >> >> > } > >> >> > > >> >> > ! if (same) > >> >> > return fold_build2_loc (loc, MULT_EXPR, type, > >> >> > fold_build2_loc (loc, code, type, > >> >> > fold_convert_loc (loc, type, alt0), > >> >> > fold_convert_loc (loc, type, alt1)), > >> >> > fold_convert_loc (loc, type, same)); > >> >> > > >> >> > ! return NULL_TREE; > >> >> > } > >> >> > > >> >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >> >> > --- 6975,7010 ---- > >> >> > } > >> >> > } > >> >> > > >> >> > ! if (!same) > >> >> > ! return NULL_TREE; > >> >> > ! > >> >> > ! if (! INTEGRAL_TYPE_P (type) > >> >> > ! || TYPE_OVERFLOW_WRAPS (type) > >> >> > ! /* We are neither factoring zero nor minus one. */ > >> >> > ! || TREE_CODE (same) == INTEGER_CST) > >> >> > return fold_build2_loc (loc, MULT_EXPR, type, > >> >> > fold_build2_loc (loc, code, type, > >> >> > fold_convert_loc (loc, type, alt0), > >> >> > fold_convert_loc (loc, type, alt1)), > >> >> > fold_convert_loc (loc, type, same)); > >> >> > > >> >> > ! /* Same may be zero and thus the operation 'code' may overflow. Likewise > >> >> > ! same may be minus one and thus the multiplication may overflow. Perform > >> >> > ! the operations in an unsigned type. */ > >> >> > ! tree utype = unsigned_type_for (type); > >> >> > ! tree tem = fold_build2_loc (loc, code, utype, > >> >> > ! fold_convert_loc (loc, utype, alt0), > >> >> > ! fold_convert_loc (loc, utype, alt1)); > >> >> > ! /* If the sum evaluated to a constant that is not -INF the multiplication > >> >> > ! cannot overflow. */ > >> >> > ! if (TREE_CODE (tem) == INTEGER_CST > >> >> > ! && ! wi::eq_p (tem, wi::min_value (TYPE_PRECISION (utype), SIGNED))) > >> >> > ! return fold_build2_loc (loc, MULT_EXPR, type, > >> >> > ! fold_convert (type, tem), same); > >> >> > ! > >> >> > ! return fold_convert_loc (loc, type, > >> >> > ! fold_build2_loc (loc, MULT_EXPR, utype, tem, > >> >> > ! fold_convert_loc (loc, utype, same))); > >> >> > } > >> >> > > >> >> > /* Subroutine of native_encode_expr. Encode the INTEGER_CST > >> >> > >> >> Sorry if you already know, but this part means that we can no longer > >> >> vectorise: > >> >> > >> >> int > >> >> f (int *x, int b1, int b2, int b3) > >> >> { > >> >> int foo = 0; > >> >> for (int i1 = 0; i1 < b1; ++i1) > >> >> for (int i2 = 0; i2 < b2; ++i2) > >> >> for (int i3 = 0; i3 < b3; ++i3) > >> >> foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > >> >> return foo; > >> >> } > >> >> > >> >> We now convert all the arithmetic in the [...] to unsigned int > >> >> and reassociate it so that the "- 1" is applied last. We then assume > >> >> that the overflow in this subtraction is well-defined and that the > >> >> &x[...] might not be linear for the inner loop. > >> > > >> > Can you file a PR? > >> > >> OK, filed as PR81082. > >> > >> > # i3_44 = PHI <i3_32(6), 0(4)> > >> > i3.2_7 = (unsigned int) i3_44; > >> > _9 = i3.2_7 + _34; > >> > _10 = (int) _9; > >> > _11 = (long unsigned int) _10; > >> > _12 = _11 * 4; > >> > _13 = x_30(D) + _12; > >> > _14 = *_13; > >> > ... > >> > i3_32 = i3_44 + 1; > >> > > >> > so we have *(x_30(D) + 4 * ((sizetype)(int)((unsigned){ 0, + , 1} + X)) > >> > > >> > It might mean that we need to avoid this re-association. Not sure how > >> > to detect worthwhile vs. not worthwhile cases during folding. At least > >> > I see no easy way to recover from it in SCEV analysis unless the > >> > number of iterations is constrained more than in your example. > >> > >> Yeah, in the example that this was reduced from, not reassociating > >> is far more preferable to changing the types. But like you say, > >> I've no idea how we'd know that at this stage. > > > > In the past I played with not obfuscating things during FE lowering > > so early, namely not expose the * 4 but keep the original types of > > array / pointer indexes. On the original mem-ref branch I had > > a IDX_EXPR that allowed to chain those (a widen-multiply-accumulate > > op). > > > > That said, the context where this association is not interesting is > > address context and the multiplication to the byte offset. > > > > Of course in your case there's also b3 factored out which looks > > like a generally interesting transform (but eventually harmful in address > > context again). > > > > From the FE we get > > > > *(x + (sizetype) ((long unsigned int) (((i1 * b2) * b3 + i2 * b3) + (i3 + > > -1)) * 4)) > > > > and SCEV uses the fact that the mults/adds have undefined behavior > > on overflow. The same missed optimization occurs if you make all those > > vars unsigned (with or without the folding in question): > But with unsigned type it's not missed optimization because > computation could overflow and result in non-scev address. Correct. > In this > case the loop needs to be versioned under overflow/non-overflow to be > parallelized/vectorized, right? Or split for cases we can easily > infer boundary condition of overflow. Yes. Richard. > Thanks, > bin > > > > #define U unsigned > > int > > f (int *x, U int b1, U int b2, U int b3) > > { > > int foo = 0; > > for (U int i1 = 0; i1 < b1; ++i1) > > for (U int i2 = 0; i2 < b2; ++i2) > > for (U int i3 = 0; i3 < b3; ++i3) > > foo += x[i1 * b2 * b3 + i2 * b3 + (i3 - 1)]; > > return foo; > > } > > > > It works again for unsigned long as there can be no valid object > > where the computation could wrap around. > > > > Richard. > > > > -- > > Richard Biener <rguenther@suse.de> > > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg) > >
Index: gcc/tree-tailcall.c =================================================================== --- gcc/tree-tailcall.c (revision 248722) +++ gcc/tree-tailcall.c (working copy) @@ -184,7 +184,8 @@ suitable_for_tail_call_opt_p (void) containing the value of EXPR at GSI. */ static tree -independent_of_stmt_p (tree expr, gimple *at, gimple_stmt_iterator gsi) +independent_of_stmt_p (tree expr, gimple *at, gimple_stmt_iterator gsi, + bitmap to_move) { basic_block bb, call_bb, at_bb; edge e; @@ -196,6 +197,9 @@ independent_of_stmt_p (tree expr, gimple if (TREE_CODE (expr) != SSA_NAME) return NULL_TREE; + if (bitmap_bit_p (to_move, SSA_NAME_VERSION (expr))) + return expr; + /* Mark the blocks in the chain leading to the end. */ at_bb = gimple_bb (at); call_bb = gimple_bb (gsi_stmt (gsi)); @@ -250,13 +254,16 @@ independent_of_stmt_p (tree expr, gimple return expr; } +enum par { FAIL, OK, TRY_MOVE }; + /* Simulates the effect of an assignment STMT on the return value of the tail recursive CALL passed in ASS_VAR. M and A are the multiplicative and the additive factor for the real return value. */ -static bool -process_assignment (gassign *stmt, gimple_stmt_iterator call, tree *m, - tree *a, tree *ass_var) +static par +process_assignment (gassign *stmt, + gimple_stmt_iterator call, tree *m, + tree *a, tree *ass_var, bitmap to_move) { tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE; tree dest = gimple_assign_lhs (stmt); @@ -276,7 +283,7 @@ process_assignment (gassign *stmt, gimpl if (gimple_assign_cast_p (stmt)) { if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var))) - return false; + return FAIL; /* Even if the type modes are the same, if the precision of the type is smaller than mode's precision, @@ -284,11 +291,11 @@ process_assignment (gassign *stmt, gimpl if (INTEGRAL_TYPE_P (TREE_TYPE (dest)) && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest))) > TYPE_PRECISION (TREE_TYPE (dest)))) - return false; + return FAIL; } *ass_var = dest; - return true; + return OK; } switch (rhs_class) @@ -303,7 +310,7 @@ process_assignment (gassign *stmt, gimpl break; default: - return false; + return FAIL; } /* Accumulator optimizations will reverse the order of operations. @@ -311,42 +318,45 @@ process_assignment (gassign *stmt, gimpl that addition and multiplication are associative. */ if (!flag_associative_math) if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl)))) - return false; + return FAIL; - if (rhs_class == GIMPLE_UNARY_RHS) + if (rhs_class == GIMPLE_UNARY_RHS + && op0 == *ass_var) ; else if (op0 == *ass_var - && (non_ass_var = independent_of_stmt_p (op1, stmt, call))) + && (non_ass_var = independent_of_stmt_p (op1, stmt, call, + to_move))) ; else if (op1 == *ass_var - && (non_ass_var = independent_of_stmt_p (op0, stmt, call))) + && (non_ass_var = independent_of_stmt_p (op0, stmt, call, + to_move))) ; else - return false; + return TRY_MOVE; switch (code) { case PLUS_EXPR: *a = non_ass_var; *ass_var = dest; - return true; + return OK; case POINTER_PLUS_EXPR: if (op0 != *ass_var) - return false; + return FAIL; *a = non_ass_var; *ass_var = dest; - return true; + return OK; case MULT_EXPR: *m = non_ass_var; *ass_var = dest; - return true; + return OK; case NEGATE_EXPR: *m = build_minus_one_cst (TREE_TYPE (op0)); *ass_var = dest; - return true; + return OK; case MINUS_EXPR: if (*ass_var == op0) @@ -358,12 +368,10 @@ process_assignment (gassign *stmt, gimpl } *ass_var = dest; - return true; - - /* TODO -- Handle POINTER_PLUS_EXPR. */ + return OK; default: - return false; + return FAIL; } } @@ -523,6 +531,7 @@ find_tail_calls (basic_block bb, struct since we are running after dce. */ m = NULL_TREE; a = NULL_TREE; + auto_bitmap to_move; abb = bb; agsi = gsi; @@ -540,27 +549,36 @@ find_tail_calls (basic_block bb, struct } stmt = gsi_stmt (agsi); - - if (gimple_code (stmt) == GIMPLE_LABEL - || gimple_code (stmt) == GIMPLE_NOP) - continue; - if (gimple_code (stmt) == GIMPLE_RETURN) break; - if (gimple_clobber_p (stmt)) - continue; - - if (is_gimple_debug (stmt)) + if (gimple_code (stmt) == GIMPLE_LABEL + || gimple_code (stmt) == GIMPLE_NOP + || gimple_clobber_p (stmt) + || is_gimple_debug (stmt)) continue; if (gimple_code (stmt) != GIMPLE_ASSIGN) return; /* This is a gimple assign. */ - if (! process_assignment (as_a <gassign *> (stmt), gsi, &tmp_m, - &tmp_a, &ass_var)) + par ret = process_assignment (as_a <gassign *> (stmt), gsi, + &tmp_m, &tmp_a, &ass_var, to_move); + if (ret == FAIL) return; + else if (ret == TRY_MOVE) + { + if (! tail_recursion) + return; + for (unsigned opno = 1; opno < gimple_num_ops (stmt); ++opno) + { + tree op = gimple_op (stmt, opno); + if (independent_of_stmt_p (op, stmt, gsi, to_move) != op) + return; + } + bitmap_set_bit (to_move, SSA_NAME_VERSION (gimple_assign_lhs (stmt))); + continue; + } if (tmp_a) { @@ -601,6 +619,19 @@ find_tail_calls (basic_block bb, struct if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl)))) return; + /* Move queued defs. */ + if (tail_recursion) + { + bitmap_iterator bi; + unsigned i; + EXECUTE_IF_SET_IN_BITMAP (to_move, 0, i, bi) + { + stmt = SSA_NAME_DEF_STMT (ssa_name (i)); + gimple_stmt_iterator mgsi = gsi_for_stmt (stmt); + gsi_move_before (&mgsi, &gsi); + } + } + nw = XNEW (struct tailcall); nw->call_gsi = gsi;